System and method for dynamic noise adaptation for robust automatic speech recognition

Abstract

A speech processing method and arrangement are described. A dynamic noise adaptation (DNA) model characterizes a speech input reflecting effects of background noise. A null noise DNA model characterizes the speech input based on reflecting a null noise mismatch condition. A DNA interaction model performs Bayesian model selection and re-weighting of the DNA model and the null noise DNA model to realize a modified DNA model characterizing the speech input for automatic speech recognition and compensating for noise to a varying degree depending on relative probabilities of the DNA model and the null noise DNA model.

What is claimed is: 
     
       1. A method comprising:
 characterizing, by a computing device, a speech input based on a dynamic noise adaptation (DNA) model reflecting effects of background noise; 
 characterizing the speech input based on a null noise DNA model reflecting a null noise mismatch condition; 
 performing Bayesian model averaging using a first weighting of the DNA model and a second weighting of the null noise DNA model; 
 re-weighting the DNA model and the null noise DNA model by adjusting the first weighting to increase a probability of the DNA model predicted by the Bayesian model averaging when the DNA model is more likely than the null noise DNA model to best characterize the speech input; and 
 performing recognition of the speech input using the re-weighted DNA model and the re-weighted null noise DNA model. 
 
     
     
       2. The method of  claim 1 , comprising re-weighting the DNA model and the null noise DNA model by adjusting the first weighting to decrease the probability of the DNA model predicted by the Bayesian model averaging when the DNA model is less likely than the null noise DNA model to best characterize the speech input. 
     
     
       3. The method of  claim 2 , comprising re-weighting the DNA model and the null noise DNA model by adjusting the first weighting to decrease the probability of the DNA model predicted by the Bayesian model averaging to be zero when the DNA model is less likely than the null noise DNA model to best characterize the speech input. 
     
     
       4. The method of  claim 1 , comprising re-weighting the DNA model and the null noise DNA model by adjusting the second weighting to decrease the probability of the null noise DNA model to be zero when the DNA model is more likely than the null noise DNA model to best characterize the speech input. 
     
     
       5. The method of  claim 1 , wherein the DNA model comprises a probability-based noise model reflecting transient components and evolving components of a current noise estimate of the background noise. 
     
     
       6. The method of  claim 1 , comprising:
 modeling a transient component of a noise process of the speech input at each frequency band as zero mean and Gaussian; 
 modeling a channel distortion of the speech input as a stochastically adapted parameter; 
 approximating a noise posterior at each given frame of the speech input as Gaussian; and 
 iteratively estimating a conditional posterior of a level of the background noise and a speech of the speech input for each speech Gaussian. 
 
     
     
       7. A system comprising:
 at least one processor; and 
 one or more non-transitory computer-readable media storing executable instructions that, when executed by the at least one processor, cause the system to:
 characterize a speech input based on a dynamic noise adaptation (DNA) model reflecting effects of background noise; 
 characterize the speech input based on a null noise DNA model reflecting a null noise mismatch condition; 
 perform Bayesian model averaging using a first weighting of the DNA model and a second weighting of the null noise DNA model; 
 re-weight the DNA model and the null noise DNA model by adjusting the first weighting to increase a probability of the DNA model predicted by the Bayesian model averaging when the DNA model is more likely than the null noise DNA model to best characterize the speech input; and 
 perform recognition of the speech input using the re-weighted DNA model and the re-weighted null noise DNA model. 
 
 
     
     
       8. The system of  claim 7 , wherein the one or more non-transitory computer-readable media store executable instructions that, when executed by the at least one processor, cause the system to:
 re-weight the DNA model and the null noise DNA model by adjusting the first weighting to decrease the probability of the DNA model predicted by the Bayesian model averaging when the DNA model is less likely than the null noise DNA model to best characterize the speech input. 
 
     
     
       9. The system of  claim 8 , wherein the one or more non-transitory computer-readable media store executable instructions that, when executed by the at least one processor, cause the system to:
 re-weight the DNA model and the null noise DNA model by adjusting the first weighting to decrease the probability of the DNA model predicted by the Bayesian model averaging to be zero when the DNA model is less likely than the null noise DNA model to best characterize the speech input. 
 
     
     
       10. The system of  claim 7 , wherein the one or more non-transitory computer-readable media store executable instructions that, when executed by the at least one processor, cause the system to:
 re-weight the DNA model and the null noise DNA model by adjusting the second weighting of the null noise DNA model to be zero when the DNA model is more likely than the null noise DNA model to best characterize the speech input. 
 
     
     
       11. The system of  claim 7 , wherein the DNA model comprises a probability-based noise model reflecting transient components and evolving components of a current noise estimate of the background noise. 
     
     
       12. The system of  claim 7 , wherein the DNA model includes a speech model, a noise model, a channel model, and an interaction model that describes how the speech model, the noise model, and the channel model combine to generate the speech input. 
     
     
       13. The system of  claim 7 , wherein the one or more non-transitory computer-readable media store executable instructions that, when executed by the at least one processor, cause the system to:
 model a transient component of a noise process of the speech input at each frequency band as zero mean and Gaussian; 
 model a channel distortion of the speech input as a stochastically adapted parameter; 
 approximate a noise posterior at each given frame of the speech input as Gaussian; and 
 iteratively estimate a conditional posterior of a level of the background noise and a speech of the speech input for each speech Gaussian. 
 
     
     
       14. One or more non-transitory computer-readable media storing executable instructions that, when executed by a processor, cause a device to:
 characterize a speech input based on a dynamic noise adaptation (DNA) model reflecting effects of background noise; 
 characterize the speech input based on a null noise DNA model reflecting a null noise mismatch condition; 
 perform Bayesian model averaging using a first weighting of the DNA model and a second weighting of the null noise DNA model; 
 re-weight the DNA model and the null noise DNA model by adjusting the first weighting to increase a probability of the DNA model predicted by the Bayesian model averaging when the DNA model is more likely than the null noise DNA model to best characterize the speech input; and 
 perform recognition of the speech input using the re-weighted DNA model and the re-weighted null noise DNA model. 
 
     
     
       15. The one or more non-transitory computer-readable media of  claim 14 , wherein the executable instructions, when executed by the processor, cause the device to:
 re-weight the DNA model and the null noise DNA model by adjusting the first weighting to decrease the probability of the DNA model predicted by the Bayesian model averaging when the DNA model is less likely than the null noise DNA model to best characterize the speech input. 
 
     
     
       16. The one or more non-transitory computer-readable media of  claim 15 , wherein the executable instructions, when executed by the processor, cause the device to:
 re-weight the DNA model and the null noise DNA model by adjusting the first weighting to decrease the probability of the DNA model predicted by the Bayesian model averaging to be zero when the DNA model is less likely than the null noise DNA model to best characterize the speech input. 
 
     
     
       17. The one or more non-transitory computer-readable media of  claim 14 , wherein the executable instructions, when executed by the processor, cause the device to:
 re-weight the DNA model and the null noise DNA model by adjusting the second weighting to decrease the probability of the null noise DNA model to be zero when the DNA model is more likely than the null noise DNA model to best characterize the speech input. 
 
     
     
       18. The method of  claim 1 , wherein the DNA model includes a speech model, a noise model, a channel model, and an interaction model that describes how the speech model, the noise model, and the channel model combine to generate the speech input. 
     
     
       19. The one or more non-transitory computer-readable media of  claim 14 , wherein the DNA model includes a speech model, a noise model, a channel model, and an interaction model that describes how the speech model, the noise model, and the channel model combine to generate the speech input. 
     
     
       20. The one or more non-transitory computer-readable media of  claim 14 , wherein the executable instructions, when executed by the processor, cause the device to:
 model a transient component of a noise process of the speech input at each frequency band as zero mean and Gaussian; 
 model a channel distortion of the speech input as a stochastically adapted parameter; 
 approximate a noise posterior at each given frame of the speech input as Gaussian; and 
 iteratively estimate a conditional posterior of a level of the background noise and a speech of the speech input for each speech Gaussian.